Theorem
$$ \text{For every } k \in \mathbb{N}, \text{we have } R(k)\ge 2^{k/2} $$
Before we start the proof, let’s look at 3 ingredients we need:
- The probabilistic method
- The union bound
- The binomial estimate
The probabilistic method
Let $\Omega$ be the set of all possible outcomes of an experiment. Suppose that each outcome is equally likely. Let $A \subseteq \Omega$ be an event. $$ \text{If } \operatorname{Prob}(A) < 1 \text{, then } \operatorname{Prob}(A^c) > 0 $$
Union Bound
Let $\Omega$ be the set of all possible outcome of an experiment. Suppose that each outcome is equally likely. Let $A_1, \ldots, A_t \subseteq \Omega$ be some events. Then: $$ \operatorname{Prob}(A_1 \cup A_2 \cup \cdots A_t) \leq \sum_{i=1}^{t}\operatorname{Prob}(A_i) $$ Proof
Consider the sets:
\begin{align*} B_1 &= A_1\\ B_2 &= A_2 \setminus A_1\\ B_3 &= A_3 \setminus (A_1 \cup A_2)\\ &\vdots\\ B_t &= A_t \setminus (A_1 \cup \cdots \cup A_t)\\ \end{align*}
Notice that:
- $B_i \cap B_j = \varnothing, \forall i \neq j$
- $B_i \subseteq A_i, \forall i$
- $A_1 \cup \cdots \cup A_t = B_1 \cup \cdots \cup B_t$
These implies that: $$ \mathbb{P}(A_1 \cup \cdots \cup A_t) = \sum_{i=1}^{t}\mathbb{P}(B_i) \leq \sum_{i=1}^{t}\mathbb{P}(A_i) $$
The binomial estimate
$$ \text{If } 1 \leq k \leq n, \text{ then } {{n} \choose {k}} \leq \left(\frac{ne}{k}\right)^k $$
Proof
We simply use that $1+x \leq e^x$ for all $x \in \mathbb{R}$.
To see why this “fact” holds, consider the tangent line of the function $f(x) = e^x$ at $x=0$.
Use the binomial theorem
\begin{align*} \left( 1 + \frac{k}{n} \right)^k &= \sum_{i=0}^{n}\left(\frac{k}{n}\right)^i{{n} \choose {i}} &&\text{by Binomial Theorem}\\ &\geq \left( \frac{k}{n} \right)^k \cdot \binom{n}{k} && \text{when $i$ starts from $k$}\\ \end{align*}
Since we have the fact: $1+x \leq e^x$, when $x = \frac{k}{n}$, we have:
\begin{align*} \left( e^{k/n} \right)^n = e^k \geq \left( 1 + \frac{k}{n} \right)^k &\geq \left( \frac{k}{n} \right)^k \cdot \binom{n}{k} \\ \iff \left( \frac{en}{k} \right)^k &\geq \binom{n}{k} \end{align*}
NOW, lets start our proof of the “lower bound” theorem in the beginning
Proof of $R(k)\geq 2^{k/2}$
Intuition
We want to show that whenever $n$ is smaller than some number, it is possible to have some coloring of $K_n$ such that $\nexists $monochromatic $K_k$.
Approaching the proof
Let $n \in \mathbb{N}$ to be chosen later.
Experiment: color each edge of $K_n$ uniformly at random with red or blue.
Set of all possible outcome is: $$ \Omega = \set{\text{red, blue}}^{{n} \choose {2}} $$
Then, choose a $k$. Label all the copies of $K_k$ inside $K_n$. There are ${{n} \choose {k}}$ of them.
Our events will be of the following form: $$ A_i = \set{i^{th} \text{ clique is monochromatic}} $$
Consider the probability of each of these $A_i$. They should have the same probability. So $\forall i \in \set{1, \ldots, \binom{n}{k} }, A_i = A_1$. For each edge in the $K_k$ graph, it can be colored either red or blue. By applying the multiplication principle, this is: $$ \mathbb{P}(A_i) = 2 \cdot \left( \frac{1}{2} \right)^{|E(K_k)|} = 2 \cdot \left( \frac{1}{2} \right)^{\binom{k}{2}} $$
We want to know what is the probability of the event $A$, that for all coloring of $K_n$, there exists a monochromatic $K_k$:
\begin{align*} \mathbb{P}\left(A\right) &= \mathbb{P}\left(\bigcup_{i=1}^{\binom{n}{k}}A_i\right) = \mathbb{P}\left(\bigcup_{i=1}^{\binom{n}{k}}A_1\right)\\ &= \binom{n}{k}\mathbb{P}\left( A_1 \right) = \binom{n}{k} \cdot 2 \cdot \left( \frac{1}{2} \right)^{\binom{k}{2}}\\ &\leq \left( \frac{ne}{k} \right)^k \cdot 2 \cdot \left( \frac{1}{2} \right)^{\binom{k}{2}} && \text{binomial estimate}\\ &= \left( \frac{ne}{k} \right)^k \cdot 2 \cdot \left( \frac{1}{2} \right)^{k(k-1)/2}\\ &= \left( \frac{ne\cdot 2^{1/k}}{k\cdot 2^{(k-1)/2}} \right)^k \end{align*}
HERE COMES THE MAGIC!!!
Pick $n = 2^{k/2}$ for this expression, then cancel some terms.
\begin{align*} P(A) &\leq \left( \frac{2^{k/2} \cdot e\cdot 2^{1/k}}{k\cdot 2^{(k-1)/2}} \right)^k\\ &= \left( \frac{2^{1/2} \cdot e \cdot 2^{1/k}}{k} \right)^k \end{align*}
Then, we argue that this is smaller than 1 for all $k \geq 5$. To see why, just check the correctness of the inequality:
$$
\sqrt{2} \cdot e \cdot 2^{1/k} < k
$$
When $k = 5$, left side is about 4.416. When $k$ gets larger, it left side gets closer to $\sqrt{2} e$, which must be smaller than $k$. For $k < 5$ cases, we can just check them by hand.
When we have the result that $P(A) < 1$, by the probabilistic method, we know the probability of the existance of a coloring such that no monochromatic $K_k$ appears in $K_n$.
Personally, I have one confusion: when $n = 2^{k/2}$, we have the above result. But we are trying to prove $R(k)\geq 2^{k/2}$… so…🤔?
Reference:
https://lmattos.web.illinois.edu/math-413-lecture-log/, MATH 413, Leticia Dias Mattos